OpenCoder-LLM/opc-sft-stage2

OpenCoder数据集由多个子数据集组成，包括opc-sft-stage1、opc-sft-stage2、opc-annealing-corpus、opc-fineweb-code-corpus、opc-fineweb-math-corpus和refineCode-code-corpus-meta。其中，sft-stage2数据集包含四个部分：educational_instruct、evol_instruct、mceval_instruct和package_instruct。educational_instruct使用算法语料库生成（指令、代码、测试用例）三元组，并通过Python编译器验证；evol_instruct直接使用开源版本MagicCoder-Evol-Instruct-110k；mceval_instruct直接使用开源版本McEval-Instruct；package_instruct则从pydoc中提取常见接口文档，生成与Python包相关的问题。

The sft-stage2 part of the OpenCoder dataset consists of four sub-datasets: educational_instruct, evol_instruct, mceval_instruct, and package_instruct. educational_instruct generates (instruction, code, test case) triples using an algorithmic corpus and validates them through a Python compiler, including test cases to enhance code reinforcement learning signals. evol_instruct uses the open-source version MagicCoder-Evol-Instruct-110k. mceval_instruct uses the open-source version McEval-Instruct. package_instruct extracts common interface documentation from pydoc to generate Python package-related questions.